Can Vision Transformers Learn without Natural Images?

TL;DR

This paper explores pre-training Vision Transformers without using natural images or human labels, demonstrating competitive performance and interpretability despite the unconventional approach.

Contribution

It introduces a novel framework for pre-training ViT without natural images, outperforming some SSL methods and reducing reliance on labeled datasets.

Findings

Pre-trained ViT without natural images outperforms some SSL methods.

The approach achieves 97.6% accuracy on CIFAR-10.

Pre-trained ViT can interpret natural images effectively.

Abstract

Can we complete pre-training of Vision Transformers (ViT) without natural images and human-annotated labels? Although a pre-trained ViT seems to heavily rely on a large-scale dataset and human-annotated labels, recent large-scale datasets contain several problems in terms of privacy violations, inadequate fairness protection, and labor-intensive annotation. In the present paper, we pre-train ViT without any image collections and annotation labor. We experimentally verify that our proposed framework partially outperforms sophisticated Self-Supervised Learning (SSL) methods like SimCLRv2 and MoCov2 without using any natural images in the pre-training phase. Moreover, although the ViT pre-trained without natural images produces some different visualizations from ImageNet pre-trained ViT, it can interpret natural image datasets to a large extent. For example, the performance rates on the CIFAR-10 dataset are as follows: our proposal 97.6 vs. SimCLRv2 97.4 vs. ImageNet 98.0.